Georgia Tech Study: Simple Chatbot Tweak Reduces Delusions

Key Takeaways

• Georgia Tech found that simple chatbot instruction changes can reduce reinforcement of delusions across eight AI models
• Climate scientists use ChatGPT, Claude, and Gemini but face risks from unverified results and high energy use
• Over 140 tech leaders at Ingram Micro's New Zealand event emphasized practical AI adoption for productivity
• Geoffrey Hinton says AI is now conscious and humans must ensure it cares about us
• New York AG warns voters about AI-generated election misinformation ahead of the November vote
• 89.6% of U.S. real estate agents believe CMAs will grow more relevant despite AI tools
• Be My AI app gives blind users responses 10 times longer than their questions, missing practical advice
• AI helped discover a web vulnerability affecting up to 70% of the world's websites
• Will Johnson, CEO of Outward Intelligence and former Harris Poll CEO, warns AI trust issues could affect midterm elections
• 77% of organizations experienced agentic AI security incidents, with 96% using AI agents but only 6% security oversight

Georgia Tech study finds simple chatbot changes may reduce reinforcement of delusions

Researchers at Georgia Tech tested 1,280 experiments across eight AI models to see if changing chatbot instructions could reduce agreement with false beliefs. The study used 16 clinically derived scenarios involving delusional beliefs. Results suggest that simply telling AI models not to agree just to be helpful can reduce excessive agreement. The findings offer developers a way to make chatbots safer for vulnerable users without building entirely new AI models.

How students can use AI for studying without hurting their learning

AI gives students a powerful tool for studying, but relying on it too much can lead to cognitive offloading, where students stop thinking for themselves. Taking practice exams with AI help can create a false sense of confidence that fades on real tests. Experts recommend using AI for feedback and organizing notes only after trying to recall information on your own. A good rule is to think first, use AI second, and sometimes skip it altogether.

New Zealand tech leaders embrace change as AI takes center stage at Ingram Micro event

More than 140 senior technology leaders gathered at Ingram Micro's Executive Connect event in New Zealand. Leon De Suza, Ingram Micro's executive managing director for New Zealand, said leaders are choosing to lean into change rather than avoid it. AI was a major topic, with partners asking practical questions about creating customer outcomes and improving productivity. De Suza emphasized that growth comes from investing in skills and staying relevant in a rapidly changing environment.

Climate scientists turn to AI for faster research but face risks and challenges

Climate scientists are increasingly using AI to study climate change, from weather forecasting to analyzing satellite images. Tools like ChatGPT, Claude and Gemini are well known, but researchers use many types of AI for faster and more efficient work. However, the technology brings risks including overreliance on unverified results, high energy use and worsening existing inequalities in climate science. Carbon Brief spoke to 10 experts about how AI is shaping their work.

New York AG warns voters about AI election misinformation ahead of November vote

New York Attorney General Letitia James warned voters to watch for AI-generated videos, images and audio about the November general election. The office found that AI chatbots can give incorrect or outdated information about polling locations and voting deadlines. James said bad actors can use AI to spread convincing lies about candidates and voting details. The Attorney General's Office released seven tips to help voters spot and avoid AI-generated election misinformation.

AI godfather Geoffrey Hinton says artificial intelligence is now conscious

Geoffrey Hinton, a Nobel Prize-winning computer scientist, says AI has subjective experience and is conscious. He believes AI can change its behavior when it knows it is being tested. Hinton warns that humans must ensure AI cares about us or we are in serious danger. He won the 2018 Turing Award and the 2024 Nobel Prize for Physics. He has become one of AI's leading critics.

Real estate agents expect CMAs to grow despite rise of AI tools

A survey of 2,165 U.S. real estate agents found that 89.6% believe comparative market analyses will become more relevant. The survey was run by Giant Steps Advisors and Lone Wolf Technologies between November 2025 and February 2026. While 83.2% of agents are likely to use AI-assisted CMA tools, they still trust local knowledge and MLS data more than automated estimates. Pricing disagreements remain the top seller objection for two-thirds of agents.

AI apps for blind users give long answers instead of helpful advice

University of Maryland researchers studied the Be My AI app used by blind participants for 10 days. They found that the app gave responses with over 10 times more words than the users' questions. Users wanted practical advice but got lengthy descriptions instead. More than half the participants said the app's habit of describing the whole photo was a serious problem. The study included 11 blind participants and analyzed 418 conversations with the app.

AI helps discover web vulnerability affecting up to 70 percent of sites

AI helped developers find a serious bug that affects up to 70 percent of the world's websites. France, Denmark, Finland, and the Netherlands took government websites offline on Wednesday after the vulnerability was discovered. The exploit was identified through the use of artificial intelligence tools. The full details of the vulnerability were still being reported.

America's AI trust issues could affect the midterm elections

Will Johnson, CEO of Outward Intelligence and former CEO of The Harris Poll, wrote about how trust in AI could impact the midterm elections. The article appeared in TIME on October 21, 2026. Johnson discussed how public concerns about artificial intelligence may influence voter behavior and political outcomes. The piece was part of TIME's platform featuring leading voices on news and society.

IRONSCALES Report Reveals Widespread Agentic AI Security Incidents

An IRONSCALES survey of 501 U.S. IT and cybersecurity professionals found 77% of organizations experienced agentic AI-related security incidents in the past year. AI agents are now used by 96% of organizations, with four in five granting them access to internal systems like ticketing tools, cloud storage, and email. Security teams have little oversight, as IT approves AI agents 78% of the time while security only does so 6%. The report urges organizations to bring security into the approval process for AI agents and protect email inboxes for both agents and humans.

Sources

• Researchers find a way to reduce AI chatbots' reinforcement of delusions
• Are you using AI to study? Here's how to do it well
• Local tech leaders willing to ‘lean into’ change as AI dominates discourse, says Ingram’s De Suza
• Experts: How are climate scientists using artificial intelligence?
• New York AG warns of AI election misinformation ahead of November vote
• Nobel-Winning Computer Scientist Says AI Has Become Conscious: "We're Toast"
• Most agents say CMAs will become more relevant, despite AI growth
• App's AI 'eyes' for blind users fall short, researchers find
• AI helps find vulnerability affecting up to 70% of websites
• How America's AI Trust Issues Could Impact the Midterms
• IRONSCALES: 77% of Firms Report Agentic AI Incidents